Survival Modeling from Whole Slide Images Via Patch-Level Graph Clustering and Mixture Density Experts

We present a modular framework for predicting cancerspecific survival from whole-slide pathology images (WSIs). The approach integrates four main components: (i) Quantile-Based Patch Filtering, which employs quantile-based thresholding to identify prognostically informative tissue regions; (ii) Graph Regularized Patch Clustering using$k$-NN graph to model phenotype-level heterogeneity through spatial-morphological coherence; (iii) Hierarchical Feature Aggregation for learning intra- and inter-cluster dependencies; and (iv) an Expert Guided Mixture Density Modeling module to estimate complex survival distributions using Gaussian distributions. The proposed model achieves a concordance index of$0.653 \pm 0.037$on TCGA-LUAD,$0.719 \pm 0.011$on TCGA-KIRC, and$0.733 \pm 0.037$on TCGA-BRCA, surpassing current state-of-the-art methods.